machine_translation_args.md

Training

• learning rate
• dropout
• max token
• clip norm
• tokenization method
• update freq (mini-batch with delayed update)
• optimizer
• learning rate scheduler
• warmup update
• warmup init learning rate
• min learning rate
• label smoothing
• quantization (fp16)
• share all embedding
• criterion (cross entropy)

Generation

• beam search size
• max length (ax+b where x is the original length)